An Enhanced Meshfree Analysis With Intrinsic Convolutional Neural Networks

In meshfree methods,the supports of meshfree shape functions determine the interrelationship of discretized nodes,and thus the support sizes of meshfree shape functions have a significant impact on the meshfree solution accuracy.However,due to the complex formulation of meshfree shape function,currently there is still no general guidelines for meshfree support selection,and in practice mesh free support sizes are often picked up by trial and error which cannot always ensure the meshfree solution accuracy.Besides,in large deformation dynamic meshfree analysis,the nonlinear incremental computation is very timeconsuming.On the other hand,the convolutional neural network(CNN)is a machine learning approach which is quite sensitive to the spatial features.It is found that there is a strong similarity and linkage between the CNN receptive fields and meshfree supports,and accordingly a meshfree intrinsic CNN is proposed for the rational selection of meshfree supports and the efficient prediction of meshfree numerical solutions.Firstly,based on the natural similarity between the receptive fields and the meshfree supports,a meshfree intrinsic CNN framework is proposed,where the architecture design and hyperparameter selection for CNN are particularly discussed.In the context of the proposed meshfree intrinsic CNN framework,several CNN architectures are presented for steady problems in order to optimize meshfree support sizes and predict numerical solutions either sequentially or synchronously.Subsequently,by taking advantage of the internal similarity between the long short-term memory network(LSTM)and the Newmark transient analysis,a LSTM module is further introduced into meshfree intrinsic CNN to enable a fast CNN algorithm for the prediction of dynamic responses.Finally,the proposed methodology is embedded into the large deformation meshfree simulation of slope failure,which provides an efficient approach to predict the complex slope failure process.A series of steady,transient as well as large deformation slope sliding examples are presented to demonstrate the effectiveness of the proposed meshfree methodology enhanced by the intrinsic CNN.